Re: gnumach FTBFS: Was Re: Race condition in Mach/Hurd?

[Samuel Thibault]

Svante Signell, le Thu 19 May 2011 10:17:16 +0200, a écrit :
> On Thu, 2011-05-19 at 09:57 +0200, Samuel Thibault wrote:
> > Samuel Thibault, le Fri 13 May 2011 21:12:43 +0200, a écrit :
> > > This looks extremely odd.  Labels like .L430 are not supposed to be
> > > external symbols, but file-local symbols.
> > 
> > rgrep tells it's in liblinux.a_de4x5.o, objdump -x -D tells it's in
> > de4x5_alloc_rx_buff and the address. gdb's l * 0xaddress tells that it's
> > actually from skb_put, the "here" label thing.  It looks like a gcc 4.6
> > bug with labels. I've simploy dropped it, panic() triggers the debugger
> > which can dump a backtrace anyway.
> 
> Good hunting, looks like the problem was deeply hidden. Did you report
> the label bug to gcc upstream?

Ideally it should be reported, yes, but the gnumach shows that it'll be
difficult to isolate the issue, since the skb_put inline is used in a
lot of places in gnumach, and only one of them triggered the issue.

> Can this bug affect other code using labels compiled with gcc-4.6 -O2?

Probably.

Samuel